Stopping rules based on a posteriori probability for every pattern and every stage

Joe Suzuki, Toshiyasu Matsushima, Shigeichi Hirasawa, Hiroshige Inazumi

    Research output: Contribution to journalArticle

    Abstract

    The algorithms for pattern recognition systems considering the observation cost from the viewpoint of the system, can be reduced to the problems of stopping rule in which we must determine when the observation should be stopped. As the stopping rule in the Bayesian pattern recognition, methods based on DP search first given by K. S. Fu et al. is known to be the optimum in the sense of the minimum risk. A problem in those methods is that an exponential amount of data must be stored in memory for each observation stage, which is hard to realize. This paper discusses the new class of stopping rules, where the optimum thresholds based on a posteriori probability are chosen for each pattern and for each stage. The proposed method is not applicable to the problem where the patterns have strong correlations in the observed values along each dimension, but it can realize almost the same performance from the viewpoint of the minimum risk, while eliminating the number of parameters to be stored. The method also provides a model for the conventional method to set the border for the likelihood ratio when the a priori probabilities of the patterns are distributed uniformly.

    Original languageEnglish
    Pages (from-to)71-82
    Number of pages12
    JournalElectronics and Communications in Japan, Part III: Fundamental Electronic Science (English translation of Denshi Tsushin Gakkai Ronbunshi)
    Volume72
    Issue number8
    Publication statusPublished - 1989 Aug

    Fingerprint

    Pattern recognition systems
    Pattern recognition
    Data storage equipment
    Costs

    ASJC Scopus subject areas

    • Electrical and Electronic Engineering

    Cite this

    @article{ce64f2b6318046c6b24e38279c25c475,
    title = "Stopping rules based on a posteriori probability for every pattern and every stage",
    abstract = "The algorithms for pattern recognition systems considering the observation cost from the viewpoint of the system, can be reduced to the problems of stopping rule in which we must determine when the observation should be stopped. As the stopping rule in the Bayesian pattern recognition, methods based on DP search first given by K. S. Fu et al. is known to be the optimum in the sense of the minimum risk. A problem in those methods is that an exponential amount of data must be stored in memory for each observation stage, which is hard to realize. This paper discusses the new class of stopping rules, where the optimum thresholds based on a posteriori probability are chosen for each pattern and for each stage. The proposed method is not applicable to the problem where the patterns have strong correlations in the observed values along each dimension, but it can realize almost the same performance from the viewpoint of the minimum risk, while eliminating the number of parameters to be stored. The method also provides a model for the conventional method to set the border for the likelihood ratio when the a priori probabilities of the patterns are distributed uniformly.",
    author = "Joe Suzuki and Toshiyasu Matsushima and Shigeichi Hirasawa and Hiroshige Inazumi",
    year = "1989",
    month = "8",
    language = "English",
    volume = "72",
    pages = "71--82",
    journal = "Electronics and Communications in Japan, Part III: Fundamental Electronic Science (English translation of Denshi Tsushin Gakkai Ronbunshi)",
    issn = "1042-0967",
    publisher = "John Wiley and Sons Inc.",
    number = "8",

    }

    TY - JOUR

    T1 - Stopping rules based on a posteriori probability for every pattern and every stage

    AU - Suzuki, Joe

    AU - Matsushima, Toshiyasu

    AU - Hirasawa, Shigeichi

    AU - Inazumi, Hiroshige

    PY - 1989/8

    Y1 - 1989/8

    N2 - The algorithms for pattern recognition systems considering the observation cost from the viewpoint of the system, can be reduced to the problems of stopping rule in which we must determine when the observation should be stopped. As the stopping rule in the Bayesian pattern recognition, methods based on DP search first given by K. S. Fu et al. is known to be the optimum in the sense of the minimum risk. A problem in those methods is that an exponential amount of data must be stored in memory for each observation stage, which is hard to realize. This paper discusses the new class of stopping rules, where the optimum thresholds based on a posteriori probability are chosen for each pattern and for each stage. The proposed method is not applicable to the problem where the patterns have strong correlations in the observed values along each dimension, but it can realize almost the same performance from the viewpoint of the minimum risk, while eliminating the number of parameters to be stored. The method also provides a model for the conventional method to set the border for the likelihood ratio when the a priori probabilities of the patterns are distributed uniformly.

    AB - The algorithms for pattern recognition systems considering the observation cost from the viewpoint of the system, can be reduced to the problems of stopping rule in which we must determine when the observation should be stopped. As the stopping rule in the Bayesian pattern recognition, methods based on DP search first given by K. S. Fu et al. is known to be the optimum in the sense of the minimum risk. A problem in those methods is that an exponential amount of data must be stored in memory for each observation stage, which is hard to realize. This paper discusses the new class of stopping rules, where the optimum thresholds based on a posteriori probability are chosen for each pattern and for each stage. The proposed method is not applicable to the problem where the patterns have strong correlations in the observed values along each dimension, but it can realize almost the same performance from the viewpoint of the minimum risk, while eliminating the number of parameters to be stored. The method also provides a model for the conventional method to set the border for the likelihood ratio when the a priori probabilities of the patterns are distributed uniformly.

    UR - http://www.scopus.com/inward/record.url?scp=0024718203&partnerID=8YFLogxK

    UR - http://www.scopus.com/inward/citedby.url?scp=0024718203&partnerID=8YFLogxK

    M3 - Article

    AN - SCOPUS:0024718203

    VL - 72

    SP - 71

    EP - 82

    JO - Electronics and Communications in Japan, Part III: Fundamental Electronic Science (English translation of Denshi Tsushin Gakkai Ronbunshi)

    JF - Electronics and Communications in Japan, Part III: Fundamental Electronic Science (English translation of Denshi Tsushin Gakkai Ronbunshi)

    SN - 1042-0967

    IS - 8

    ER -